chore: replace vllm-router fork with upstream release

[mikasenghaas]

Summary

• Consume vllm-router from PyPI (vllm-project/router, >=0.1.15) instead of the pinned wheel from our PrimeIntellect-ai/router fork (v0.1.26).
• Drop the [tool.uv.sources] wheel URLs and the exclude-newer-package exemption; the dependency now resolves through the normal registry path.
• Update docs/inference.md to stop calling the router backend "our fork".
• Shut down the router cleanly on Ctrl-C: the router now runs in its own session, so a terminal SIGINT only reaches the server — the engine drains first, then inference_local terminates the router. Previously the router died instantly with the foreground group and the watcher thread logged a spurious ERROR Router exited with code 0 - shutting down plus a redundant SIGTERM. The watcher also logs a zero exit as a warning now (an external shutdown racing ours, not a crash).

Upstream 0.1.15 covers everything prime-rl invokes: the consistent_hash policy, --intra-node-data-parallel-size, --vllm-pd-disaggregation, --request-id-headers, --worker-startup-timeout-secs, --prometheus-port, and a typed /inference/v1/generate route. Extra request fields (e.g. return_routed_experts) pass through via a serde-flatten catch-all, and response bytes are forwarded verbatim, so single-replica router replay keeps working.

Breaking

Fork-only router features are gone with this swap:

• Routed-experts merge for P/D disaggregation: upstream does not merge prefill/decode routed-experts payloads. Router replay (trainer.enable_router_replay) with a disaggregated deployment will not return merged expert traces. Single-replica replay is unaffected.
• JWT auth and per-run usage metrics: platform-side features of the fork; not used by prime-rl entrypoints. Platform deployments that need them should keep deploying the fork binary independently of prime-rl's dependency.

Verification

Ran the reverse-text RL example end-to-end from this branch on 2 GPUs:

uv run rl @ examples/basic/reverse-text/rl.toml --no-wandb --output-dir outputs/router-smoke

• Router started from the upstream wheel: vllm_router_rs::server: Starting router on 0.0.0.0:8000 | mode: Regular { worker_urls: ["http://localhost:8100"] } | policy: ConsistentHash.
• All rollouts flowed through the router's /inference/v1/generate route: 20/20 steps, 128/128 trainable per step, 0.0% errors, reward 0.15 → 0.70.
• Clean shutdown, final checkpoint written at step_20.

Ctrl-C shutdown, verified with a standalone uv run inference --vllm.model Qwen/Qwen3-0.6B and SIGINT to its foreground process group:

• Before: router died with the group; watcher logged ERROR Router exited with code 0 - shutting down and re-signalled the process.
• After: engine drains first, then the router logs Received terminate signal, starting graceful shutdown; no ERROR lines, no leftover processes.